Feature Genes Selection Study In Gene Expression Profile-based Tumor Classification

Up to the present, it is still very difficult to cure the malignant tumor because of its many subtypes. Early diagnosis and correct classification are helpful to the clinical therapy of the patients, and therefore to improving the quality of their lives. Current technology for tumor diagnosis and classification is based on the histopathological features of patients. However, the patients with similar histopathological features often have significant difference in clinical therapy and prognosis.Gene chip is a high-throughput technology, which can be used to monitor expression levels of tens of thousands genes at the same time. We usually call the expression levels of many genes across a certain number of samples as the gene expression profile. Current studies indicate that, even through some tumors have the similar histopathological features, there maybe exist major differences in their gene expression profiles, which provide a possibility to accurately predict tumor type with similar histopathologic features using gene expression profile. Up to now, the application of gene expression profile in tumor classification has been extensively explored. Developing an easy and convenience tool for tumor classification and diagnosis is undoubtedly to bring large benefits to patients and society.Gene expression profile-based sample classification contains two important parts:features gene selection and classifier construction.Due to the high cost of gene-chip and availability of samples, gene expression profile often only contains a small number of samples depicted by expression levels of tens of thousands of genes.In addition, there are many redundant information due to the strong correlation of genes. It is a typical problem with high dimension of features and high noise. Therefore, feature selection plays an important role in gene expression profile-based sample classification. Based on the wrapper method GA/KNN, which has been successfully applied in many cases, here a new feature selection scheme-GA/WV, has been presented, which is a combination of GA and weight voting classifier. Finally, the better performance of GA/WV than T tests has been demonstrated by the following gene expression datasets of Gloub and Armstrong, a protein expression profile from our Lab.